You are on page 1of 36

Distance Methods

Focus is on estimation of density


( d abundance)
(and b d ) from
f closed
l d
populations
►D =N/A

► Often we will directly estimate D and infer N


ffrom kknowledge
l d off A
Methods we won’t
won t cover
► Point
Point--to
to--object
► Nearest neighbor
► Etc
Et
► Typically
used in plant studies
► Assume either
ƒ Restrictive assumptions about spatial
distribution to estimate D
ƒ D is known–
known– estimating g pattern
p
Line transect
► One or more lines traversed by observer

► Count individuals (or groups) detected from


line and measure distance from line to
detected group

► Use counts and distance to estimate


ƒ Detection functions
ƒ Density
Line Transect
► Distance
► Angle
Key assumptions
► Random sampling
► Animals directly on the transect line (point)
detected with 100% probability
► Animals do not move in response to
observer in a way that permits double
counting
► Distances and angles measured accurately
► Sufficient number of detections to permit
estimation ( 60-
60-80)
In a perfect world

All individuals assumed detected


Real world-
world- a fraction are not
d t t d
detected

Uniform Half normal


Some math
► Ifdetection is perfect then estimation is
easy

► Our problem
problem–– determine an “effective area”
=(L*2*half-
=(L*2*half
(L*2*h lf-width)
idth) off detection
d t ti as an
adjustment
More formally
x= (perp)
( ) distance
di t off an individual
i di id l presentt from
f line
li

f1(x)
(x)=distribution
distribution of distances for all individuals present

y= 1 if individual detected=0 if not detected

f2(y)=distribution of detections for all individuals present

f(x|y=1)=f(x)=distribution of distances given animal is detected

f(y=1|x)=g(x)=probability of detection at distance x


By Bayes’
Bayes Theorem

g ( x) Detection function of distance


f ( x) = w
Wh t we observe
What b

o
g ( w)

When x=0 (on the line) But by assumption g(0)=1, so

g ( 0) 1 1
f ( 0) = f ( 0) = =
w w
μ
∫ g (w)
o
∫ g (w)
o
To get density

ˆf (0) = 1
μ̂
μ

So

Estimate from data


n ˆ
f ( 0 )
D=
ˆ
Known
2L
T gett variance
To i given
i
estimates of f(x)

(We’ll come back to f(x) later…)


Variance

detectability
Population (counts)
Ways to estimate variance
► Empiricalestimates by replicate transects
(separate estimates of D)

► Jackknife estimator

► Estimate components separately, then combine


► If no replicate transects
ƒ Poisson or other model assumptions
Program DISTANCE
Program Distance
► Key function + series adjustment

► GOF

► Model selection via AIC


Uniform Negative exponential

Half-normal Hazard
Goodness of fit

k
(ni − nπˆ i ) 2
χ =∑
2

i =1 nπˆ i
Confidence intervals
► Normal approximation

► Log
L transformation
t f ti

► Profile likelihood
ƒ Not full likelihood methods
Sampling design

► Random

► Stratified random
Sample size
► Sampling effort
ƒ Pilot study

► If n encounters not sufficient to


estimate D
► Can also calculate CV at a
specified
p maximum effort
1/ 2
⎡ 3 ⎤
CVmin ( Dˆ ) = ⎢ ⎥
⎣ Lmax (n0 / Lo ) ⎦
Example (1)
► Initial
sample of effort 10km
► CV of estimated D =0.2
► Desired
D i d CV=0.1
CV 0 1
► Need total effort of 40km (30km more)

2
⎡ CV ( D) ⎤
ˆ ⎡ 0.2 ⎤
2

L=⎢ ⎥ L0 = ⎢ ⎥ 10 = 40
⎣ CV 0 ⎦ ⎣ 0. 10 ⎦
Example (2)
► Initial
sample of effort 1 km
► Detect n=25
► Desired
D i d CV=0.1
CV 0 1
► Need total effort of 12km (11km more)

3 L0 3 1
L= 2
= 2
= 12
CV0 n0 0.10 25
Can also calculate CV at a
specified
ifi d maximum
i effort
ff t
E
E.g., mostt effort
ff t possible
ibl iis 5k
5km

1/ 2 1/ 2
⎡ 3 ⎤ ⎡ 3 ⎤
CVmin ( Dˆ ) = ⎢ ⎥ =⎢ ⎥ = 0.15
⎣ Lmax (n0 / Lo ) ⎦ ⎣ 5(25 / 1) ⎦
Sample size
► Can also consider the costs of sampling (money,
time, labor,etc.)
ƒ L line length vs n replicate lines

Cx = costs
Stratified random sampling
► Lines
es ass
assigned
g ed randomly
a do y to st
strata
ata

► Density
y estimation
ƒ Weighted average across strata
► Detection estimation
ƒ Pooled
ƒ Stratum specific (AIC or LRT)
► Allocation
ƒ Proportional
ƒ Optimal
p
Experimental design
Randomized to T/C, random within
Components of variance
► Estimates of D vary from experimental unit to unit
because
ƒ Actual variation in D (among
g units, treatment)
ƒ Sample variation in estimates of D
► Need to separate these
ƒ Biological interest is on variation in actual D
ƒ Implication for study design
► Variation
i i in
i D – more replication
li i off unitsi
► Sample variation–
variation– more effort at each unit
► We’ll
We ll consider this in detail in lab
Point sampling
Point sampling
► Same basic theory as line transect

► Distances
Di t radial
di l distances
di t from
f point
i t

► Assumption of h(0) (perfect detection at


zero distance))
Sample size
► Effort now k number of sample points

► Initial
I iti l k,
k n

3 k0 3 5
k= 2
= 2
= 60
CV0 n0 0.10 25